Good article. I think another key aspect is computational irreducibility. That is to say dynamics for which computation cannot move you forward in time (in sense that you can't work out how things will be in advance)

For example certain fluid mechanics are generally considered computationally irreducible because simulating the behaviour of fluids requires tracing each step of their evolution, making it impossible to predict their future state without explicitly computing each time step (as in second for second). So basically in these cases the only way to know how it'll work out in our physics-governed world is to wait and see. (There's also an argument to be made that everything is computationally irreducible at the end of the day, but that's another thing.)

If you take this aspect and apply it to boring old jobs, there are many things in the world of work that are quite obvioulsy computationally irreducible. Meaning that an AI can't really speed things up. A business deal depends on what she says, then what he says, then if it rains here, then if it's sunny there... each step being part of a computationally irreducible chain. (Business is fluid, as they say.)

This is what makes me most skeptical of the AI productivity boost.

Good video here: